MULTICORE WITH L2 WORKING
This commit is contained in:
@@ -1,5 +1,5 @@
|
||||
`include "VX_cache_config.v"
|
||||
|
||||
`include "VX_define.v"
|
||||
module VX_bank
|
||||
#(
|
||||
// Size of cache in bytes
|
||||
@@ -54,13 +54,13 @@ module VX_bank
|
||||
input wire delay_req,
|
||||
input wire [NUMBER_REQUESTS-1:0] bank_valids,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] bank_addr,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] bank_writedata,
|
||||
input wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] bank_writedata,
|
||||
input wire [4:0] bank_rd,
|
||||
input wire [1:0] bank_wb,
|
||||
input wire [NUMBER_REQUESTS-1:0][1:0] bank_wb,
|
||||
input wire [31:0] bank_pc,
|
||||
input wire [`NW_M1:0] bank_warp_num,
|
||||
input wire [2:0] bank_mem_read,
|
||||
input wire [2:0] bank_mem_write,
|
||||
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_read,
|
||||
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_write,
|
||||
output wire reqq_full,
|
||||
|
||||
// Output Core WB
|
||||
@@ -70,8 +70,9 @@ module VX_bank
|
||||
output wire [4:0] bank_wb_rd,
|
||||
output wire [1:0] bank_wb_wb,
|
||||
output wire [`NW_M1:0] bank_wb_warp_num,
|
||||
output wire [31:0] bank_wb_data,
|
||||
output wire [`WORD_SIZE_RNG] bank_wb_data,
|
||||
output wire [31:0] bank_wb_pc,
|
||||
output wire [31:0] bank_wb_address,
|
||||
|
||||
// Dram Fill Requests
|
||||
output wire dram_fill_req,
|
||||
@@ -83,25 +84,18 @@ module VX_bank
|
||||
// Dram Fill Response
|
||||
input wire dram_fill_rsp,
|
||||
input wire [31:0] dram_fill_addr,
|
||||
input wire[`BANK_LINE_SIZE_RNG][31:0] dram_fill_rsp_data,
|
||||
input wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dram_fill_rsp_data,
|
||||
output wire dram_fill_accept,
|
||||
|
||||
// Dram WB Requests
|
||||
input wire dram_wb_queue_pop,
|
||||
output wire dram_wb_req,
|
||||
output wire[31:0] dram_wb_req_addr,
|
||||
output wire[`BANK_LINE_SIZE_RNG][31:0] dram_wb_req_data,
|
||||
output wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dram_wb_req_data,
|
||||
|
||||
// Snp Request
|
||||
input wire snp_req,
|
||||
input wire[31:0] snp_req_addr,
|
||||
|
||||
// Lower Level Cache Response
|
||||
input wire llvq_pop,
|
||||
output wire llvq_valid,
|
||||
output wire[31:0] llvq_res_addr,
|
||||
output wire[`BANK_LINE_SIZE_RNG][31:0] llvq_res_data,
|
||||
output wire[`vx_clog2(NUMBER_REQUESTS)-1:0] llvq_res_tid
|
||||
input wire[31:0] snp_req_addr
|
||||
);
|
||||
|
||||
|
||||
@@ -132,12 +126,12 @@ module VX_bank
|
||||
wire dfpq_empty;
|
||||
wire dfpq_full;
|
||||
wire[31:0] dfpq_addr_st0;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] dfpq_filldata_st0;
|
||||
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dfpq_filldata_st0;
|
||||
reg dfpq_hazard_st0;
|
||||
|
||||
assign dram_fill_accept = !dfpq_full;
|
||||
|
||||
VX_generic_queue_ll #(.DATAW(32+(`BANK_LINE_SIZE_WORDS*32)), .SIZE(DFPQ_SIZE)) dfp_queue(
|
||||
VX_generic_queue_ll #(.DATAW(32+(`BANK_LINE_SIZE_WORDS*`WORD_SIZE)), .SIZE(DFPQ_SIZE)) dfp_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.push (dram_fill_rsp),
|
||||
@@ -155,7 +149,7 @@ module VX_bank
|
||||
wire reqq_req_st0;
|
||||
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] reqq_req_tid_st0;
|
||||
wire [31:0] reqq_req_addr_st0;
|
||||
wire [31:0] reqq_req_writeword_st0;
|
||||
wire [`WORD_SIZE_RNG] reqq_req_writeword_st0;
|
||||
wire [4:0] reqq_req_rd_st0;
|
||||
wire [1:0] reqq_req_wb_st0;
|
||||
wire [`NW_M1:0] reqq_req_warp_num_st0;
|
||||
@@ -221,7 +215,7 @@ module VX_bank
|
||||
wire mrvq_valid_st0;
|
||||
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] mrvq_tid_st0;
|
||||
wire [31:0] mrvq_addr_st0;
|
||||
wire [31:0] mrvq_writeword_st0;
|
||||
wire [`WORD_SIZE_RNG] mrvq_writeword_st0;
|
||||
wire [4:0] mrvq_rd_st0;
|
||||
wire [1:0] mrvq_wb_st0;
|
||||
wire [31:0] miss_resrv_pc_st0;
|
||||
@@ -232,7 +226,7 @@ module VX_bank
|
||||
|
||||
wire miss_add;
|
||||
wire[31:0] miss_add_addr;
|
||||
wire[31:0] miss_add_data;
|
||||
wire[`WORD_SIZE_RNG] miss_add_data;
|
||||
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] miss_add_tid;
|
||||
wire[4:0] miss_add_rd;
|
||||
wire[1:0] miss_add_wb;
|
||||
@@ -311,8 +305,8 @@ module VX_bank
|
||||
|
||||
assign dfpq_pop = !dfpq_empty && !stall_bank_pipe && !dfpq_hazard_st0;
|
||||
assign mrvq_pop = !dfpq_pop && mrvq_valid_st0 && !stall_bank_pipe && !mrvq_hazard_st0;
|
||||
assign reqq_pop = !mrvq_pop && !reqq_empty && reqq_req_st0 && !stall_bank_pipe && !is_fill_st1[0] && !(reqq_hazard_st0 || (mrvq_valid_st0 && mrvq_hazard_st0)) && !is_fill_in_pipe;
|
||||
assign snrq_pop = !reqq_pop && snrq_valid_st0 && !stall_bank_pipe && !snrq_hazard_st0;
|
||||
assign reqq_pop = !mrvq_pop && !dfpq_pop && !reqq_empty && reqq_req_st0 && !stall_bank_pipe && !is_fill_st1[0] && !(reqq_hazard_st0 || (mrvq_valid_st0 && mrvq_hazard_st0)) && !is_fill_in_pipe;
|
||||
assign snrq_pop = !reqq_pop && !reqq_pop && !mrvq_pop && !dfpq_pop && snrq_valid_st0 && !stall_bank_pipe && !snrq_hazard_st0;
|
||||
|
||||
|
||||
integer st1_cycle;
|
||||
@@ -338,8 +332,8 @@ module VX_bank
|
||||
wire qual_is_fill_st0;
|
||||
wire qual_valid_st0;
|
||||
wire [31:0] qual_addr_st0;
|
||||
wire [31:0] qual_writeword_st0;
|
||||
wire [`BANK_LINE_SIZE_RNG][31:0] qual_writedata_st0;
|
||||
wire [`WORD_SIZE_RNG] qual_writeword_st0;
|
||||
wire [`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] qual_writedata_st0;
|
||||
wire [`REQ_INST_META_SIZE-1:0] qual_inst_meta_st0;
|
||||
wire qual_going_to_write_st0;
|
||||
wire qual_is_snp;
|
||||
@@ -348,14 +342,21 @@ module VX_bank
|
||||
wire valid_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire going_to_write_st1[STAGE_1_CYCLES-1:0];
|
||||
wire [31:0] addr_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire [31:0] writeword_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire [`WORD_SIZE_RNG] writeword_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire [`REQ_INST_META_SIZE-1:0] inst_meta_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire is_fill_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire [`BANK_LINE_SIZE_RNG][31:0] writedata_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire [`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] writedata_st1[STAGE_1_CYCLES-1:0];
|
||||
wire is_snp_st1 [STAGE_1_CYCLES-1:0];
|
||||
wire [31:0] pc_st1 [STAGE_1_CYCLES-1:0];
|
||||
|
||||
assign qual_is_fill_st0 = dfpq_pop;
|
||||
|
||||
// always @(*) begin
|
||||
// if (qual_is_fill_st0 && (FUNC_ID == 3)) begin
|
||||
// $display("WHAT THE FUCK FUNC_ID: %x", FUNC_ID);
|
||||
// end
|
||||
// end
|
||||
|
||||
assign qual_valid_st0 = dfpq_pop || mrvq_pop || reqq_pop || snrq_pop;
|
||||
|
||||
assign qual_addr_st0 = dfpq_pop ? dfpq_addr_st0 :
|
||||
@@ -364,11 +365,7 @@ module VX_bank
|
||||
snrq_pop ? snrq_addr_st0 :
|
||||
0;
|
||||
|
||||
assign qual_writeword_st0 = mrvq_pop ? mrvq_writeword_st0 :
|
||||
reqq_pop ? reqq_req_writeword_st0 :
|
||||
0;
|
||||
|
||||
assign qual_writedata_st0 = dfpq_pop ? dfpq_filldata_st0 : 0;
|
||||
assign qual_writedata_st0 = dfpq_pop ? dfpq_filldata_st0 : 57;
|
||||
|
||||
assign qual_inst_meta_st0 = mrvq_pop ? {mrvq_rd_st0 , mrvq_wb_st0 , mrvq_warp_num_st0 , mrvq_mem_read_st0 , mrvq_mem_write_st0 , mrvq_tid_st0 } :
|
||||
reqq_pop ? {reqq_req_rd_st0, reqq_req_wb_st0, reqq_req_warp_num_st0, reqq_req_mem_read_st0, reqq_req_mem_write_st0, reqq_req_tid_st0} :
|
||||
@@ -387,7 +384,11 @@ module VX_bank
|
||||
32'h0;
|
||||
assign qual_is_snp = snrq_pop ? 1 : 0;
|
||||
|
||||
VX_generic_register #(.N( 1 + 1 + 1 + 32 + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_SIZE_WORDS*32) + 1 + 32)) s0_1_c0 (
|
||||
assign qual_writeword_st0 = mrvq_pop ? mrvq_writeword_st0 :
|
||||
reqq_pop ? reqq_req_writeword_st0 :
|
||||
0;
|
||||
|
||||
VX_generic_register #(.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_SIZE_WORDS*`WORD_SIZE) + 1 + 32)) s0_1_c0 (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall_bank_pipe),
|
||||
@@ -399,7 +400,7 @@ module VX_bank
|
||||
genvar curr_stage;
|
||||
generate
|
||||
for (curr_stage = 1; curr_stage < STAGE_1_CYCLES; curr_stage = curr_stage + 1) begin
|
||||
VX_generic_register #(.N( 1 + 1 + 1 + 32 + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_SIZE_WORDS*32) + 1 + 32)) s0_1_cc (
|
||||
VX_generic_register #(.N( 1 + 1 + 1 + `WORD_SIZE + 32 + `REQ_INST_META_SIZE + (`BANK_LINE_SIZE_WORDS*`WORD_SIZE) + 1 + 32)) s0_1_cc (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall_bank_pipe),
|
||||
@@ -411,8 +412,8 @@ module VX_bank
|
||||
endgenerate
|
||||
|
||||
|
||||
wire[31:0] readword_st1e;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] readdata_st1e;
|
||||
wire[`WORD_SIZE_RNG] readword_st1e;
|
||||
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] readdata_st1e;
|
||||
wire[`TAG_SELECT_SIZE_RNG] readtag_st1e;
|
||||
wire miss_st1e;
|
||||
wire dirty_st1e;
|
||||
@@ -424,7 +425,7 @@ module VX_bank
|
||||
wire [`NW_M1:0] warp_num_st1e;
|
||||
wire [2:0] mem_read_st1e;
|
||||
wire [2:0] mem_write_st1e;
|
||||
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] tid_st1e;
|
||||
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] tid_st1e;
|
||||
wire fill_saw_dirty_st1e;
|
||||
wire is_snp_st1e;
|
||||
|
||||
@@ -486,9 +487,9 @@ module VX_bank
|
||||
|
||||
wire valid_st2;
|
||||
wire[31:0] addr_st2;
|
||||
wire[31:0] writeword_st2;
|
||||
wire[31:0] readword_st2;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] readdata_st2;
|
||||
wire[`WORD_SIZE_RNG] writeword_st2;
|
||||
wire[`WORD_SIZE_RNG] readword_st2;
|
||||
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] readdata_st2;
|
||||
wire miss_st2;
|
||||
wire dirty_st2;
|
||||
wire[`REQ_INST_META_SIZE-1:0] inst_meta_st2;
|
||||
@@ -498,18 +499,19 @@ module VX_bank
|
||||
wire is_snp_st2;
|
||||
wire [31:0] pc_st2;
|
||||
|
||||
VX_generic_register #(.N( 1 + 1 + 1 + 1 + 32 + 32 + 32 + (`BANK_LINE_SIZE_WORDS * 32) + 1 + 1 + `REQ_INST_META_SIZE + `TAG_SELECT_NUM_BITS + 32)) st_1e_2 (
|
||||
|
||||
VX_generic_register #(.N( 1+1+1+1+32+`WORD_SIZE+`WORD_SIZE+(`BANK_LINE_SIZE_WORDS * `WORD_SIZE) + `REQ_INST_META_SIZE + `TAG_SELECT_NUM_BITS + 32 + 2)) st_1e_2 (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall_bank_pipe),
|
||||
.flush(0),
|
||||
.in ({is_snp_st1e, fill_saw_dirty_st1e, is_fill_st1[STAGE_1_CYCLES-1], qual_valid_st1e_2, addr_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, pc_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
|
||||
.out ({is_snp_st2 , fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , pc_st2 , inst_meta_st2 })
|
||||
.in ({is_snp_st1e, fill_saw_dirty_st1e, is_fill_st1[STAGE_1_CYCLES-1] , qual_valid_st1e_2, addr_st1[STAGE_1_CYCLES-1], writeword_st1[STAGE_1_CYCLES-1], readword_st1e, readdata_st1e, readtag_st1e, miss_st1e, dirty_st1e, pc_st1e, inst_meta_st1[STAGE_1_CYCLES-1]}),
|
||||
.out ({is_snp_st2 , fill_saw_dirty_st2 , is_fill_st2 , valid_st2 , addr_st2 , writeword_st2 , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , pc_st2 , inst_meta_st2 })
|
||||
);
|
||||
|
||||
|
||||
// Enqueue to miss reserv if it's a valid miss
|
||||
assign miss_add = valid_st2 && miss_st2 && !stall_bank_pipe && !mrvq_full && !(dirty_st2 && dwbq_full);
|
||||
assign miss_add = valid_st2 && miss_st2 && !mrvq_full && !((cwbq_push && cwbq_full) || (dwbq_push && dwbq_full) || (dram_fill_req && dram_fill_req_queue_full));
|
||||
assign miss_add_pc = pc_st2;
|
||||
assign miss_add_addr = addr_st2;
|
||||
assign miss_add_data = writeword_st2;
|
||||
@@ -517,8 +519,8 @@ module VX_bank
|
||||
|
||||
|
||||
// Enqueue to CWB Queue
|
||||
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full & !llvq_full;
|
||||
wire [31:0] cwbq_data = readword_st2;
|
||||
wire cwbq_push = (valid_st2 && !miss_st2) && !cwbq_full && !((FUNC_ID == `LLFUNC_ID) && (miss_add_wb == 0));
|
||||
wire [`WORD_SIZE_RNG] cwbq_data = readword_st2;
|
||||
wire [`vx_clog2(NUMBER_REQUESTS)-1:0] cwbq_tid = miss_add_tid;
|
||||
wire [4:0] cwbq_rd = miss_add_rd;
|
||||
wire [1:0] cwbq_wb = miss_add_wb;
|
||||
@@ -528,15 +530,15 @@ module VX_bank
|
||||
wire cwbq_full;
|
||||
wire cwbq_empty;
|
||||
assign bank_wb_valid = !cwbq_empty;
|
||||
VX_generic_queue_ll #(.DATAW( `vx_clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1+1) + 32 + 32), .SIZE(CWBQ_SIZE)) cwb_queue(
|
||||
VX_generic_queue_ll #(.DATAW( `vx_clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1+1) + `WORD_SIZE + 32 + 32), .SIZE(CWBQ_SIZE)) cwb_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
.push (cwbq_push),
|
||||
.in_data ({cwbq_tid, cwbq_rd, cwbq_wb, cwbq_warp_num, cwbq_data, cwbq_pc}),
|
||||
.in_data ({cwbq_tid, cwbq_rd, cwbq_wb, cwbq_warp_num, cwbq_data, cwbq_pc, addr_st2}),
|
||||
|
||||
.pop (bank_wb_pop),
|
||||
.out_data({bank_wb_tid, bank_wb_rd, bank_wb_wb, bank_wb_warp_num, bank_wb_data, bank_wb_pc}),
|
||||
.out_data({bank_wb_tid, bank_wb_rd, bank_wb_wb, bank_wb_warp_num, bank_wb_data, bank_wb_pc, bank_wb_address}),
|
||||
.empty (cwbq_empty),
|
||||
.full (cwbq_full)
|
||||
);
|
||||
@@ -544,7 +546,7 @@ module VX_bank
|
||||
// Enqueue to DWB Queue
|
||||
wire dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && !dwbq_full && !(!fill_saw_dirty_st2 && mrvq_full);
|
||||
wire[31:0] dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_ADDR_END:0]} & `BASE_ADDR_MASK;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] dwbq_req_data = readdata_st2;
|
||||
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] dwbq_req_data = readdata_st2;
|
||||
wire dwbq_empty;
|
||||
wire dwbq_full;
|
||||
|
||||
@@ -588,7 +590,7 @@ module VX_bank
|
||||
assign dram_fill_req_addr = addr_st2 & `BASE_ADDR_MASK;
|
||||
|
||||
assign dram_wb_req = !dwbq_empty;
|
||||
VX_generic_queue_ll #(.DATAW( 32 + (`BANK_LINE_SIZE_WORDS * 32)), .SIZE(DWBQ_SIZE)) dwb_queue(
|
||||
VX_generic_queue_ll #(.DATAW( 32 + (`BANK_LINE_SIZE_WORDS * `WORD_SIZE)), .SIZE(DWBQ_SIZE)) dwb_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
|
||||
@@ -602,29 +604,8 @@ module VX_bank
|
||||
);
|
||||
|
||||
|
||||
// Lower Cache Hit
|
||||
wire llvq_empty;
|
||||
wire llvq_full;
|
||||
wire llvq_push = valid_st2 && !miss_st2 && !llvq_full && !cwbq_full;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] llvq_push_data = readdata_st2;
|
||||
wire[31:0] llvq_addr = addr_st2;
|
||||
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] llvq_tid = miss_add_tid;
|
||||
|
||||
assign llvq_valid = !llvq_empty;
|
||||
|
||||
VX_generic_queue_ll #(.DATAW(`vx_clog2(NUMBER_REQUESTS) + 32 + (`BANK_LINE_SIZE_WORDS * 32)), .SIZE(LLVQ_SIZE)) llv_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.push (llvq_push),
|
||||
.in_data ({llvq_tid , llvq_addr , llvq_push_data}),
|
||||
.pop (llvq_pop),
|
||||
.out_data({llvq_res_tid, llvq_res_addr, llvq_res_data}),
|
||||
.empty (llvq_empty),
|
||||
.full (llvq_full)
|
||||
);
|
||||
|
||||
|
||||
assign stall_bank_pipe = (cwbq_push && cwbq_full) || (llvq_push && llvq_full) || (dwbq_push && dwbq_full) || (miss_add && mrvq_full) || (dram_fill_req && dram_fill_req_queue_full);
|
||||
assign stall_bank_pipe = (cwbq_push && cwbq_full) || (dwbq_push && dwbq_full) || (miss_add && mrvq_full) || (dram_fill_req && dram_fill_req_queue_full);
|
||||
|
||||
endmodule
|
||||
|
||||
|
||||
@@ -9,13 +9,13 @@ module VX_cache
|
||||
// Number of banks {1, 2, 4, 8,...}
|
||||
parameter NUMBER_BANKS = 8,
|
||||
// Size of a word in bytes
|
||||
parameter WORD_SIZE_BYTES = 4,
|
||||
parameter WORD_SIZE_BYTES = 16,
|
||||
// Number of Word requests per cycle {1, 2, 4, 8, ...}
|
||||
parameter NUMBER_REQUESTS = 2,
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
parameter FUNC_ID = 3,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
@@ -51,15 +51,15 @@ module VX_cache
|
||||
input wire reset,
|
||||
|
||||
// Req Info
|
||||
input wire [NUMBER_REQUESTS-1:0] core_req_valid,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] core_req_addr,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] core_req_writedata,
|
||||
input wire[2:0] core_req_mem_read,
|
||||
input wire[2:0] core_req_mem_write,
|
||||
input wire [NUMBER_REQUESTS-1:0] core_req_valid,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] core_req_addr,
|
||||
input wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] core_req_writedata,
|
||||
input wire[NUMBER_REQUESTS-1:0][2:0] core_req_mem_read,
|
||||
input wire[NUMBER_REQUESTS-1:0][2:0] core_req_mem_write,
|
||||
|
||||
// Req meta
|
||||
input wire [4:0] core_req_rd,
|
||||
input wire [1:0] core_req_wb,
|
||||
input wire [NUMBER_REQUESTS-1:0][1:0] core_req_wb,
|
||||
input wire [`NW_M1:0] core_req_warp_num,
|
||||
input wire [31:0] core_req_pc,
|
||||
output wire delay_req,
|
||||
@@ -70,14 +70,15 @@ module VX_cache
|
||||
output wire [4:0] core_wb_req_rd,
|
||||
output wire [1:0] core_wb_req_wb,
|
||||
output wire [`NW_M1:0] core_wb_warp_num,
|
||||
output wire [NUMBER_REQUESTS-1:0][31:0] core_wb_readdata,
|
||||
output wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] core_wb_readdata,
|
||||
output wire [NUMBER_REQUESTS-1:0][31:0] core_wb_pc,
|
||||
output wire [NUMBER_REQUESTS-1:0][31:0] core_wb_address,
|
||||
|
||||
|
||||
// Dram Fill Response
|
||||
input wire dram_fill_rsp,
|
||||
input wire [31:0] dram_fill_rsp_addr,
|
||||
input wire [`BANK_LINE_SIZE_RNG][31:0] dram_fill_rsp_data,
|
||||
input wire [`IBANK_LINE_SIZE_RNG][31:0] dram_fill_rsp_data,
|
||||
output wire dram_fill_accept,
|
||||
|
||||
// Dram request
|
||||
@@ -86,20 +87,14 @@ module VX_cache
|
||||
output wire dram_req_read,
|
||||
output wire [31:0] dram_req_addr,
|
||||
output wire [31:0] dram_req_size,
|
||||
output wire [`BANK_LINE_SIZE_RNG][31:0] dram_req_data,
|
||||
output wire [`IBANK_LINE_SIZE_RNG][31:0] dram_req_data,
|
||||
output wire dram_req_because_of_wb,
|
||||
output wire dram_snp_full,
|
||||
|
||||
|
||||
// Snoop Req
|
||||
input wire snp_req,
|
||||
input wire[31:0] snp_req_addr,
|
||||
|
||||
// Lower Level Cache
|
||||
input wire llvq_pop,
|
||||
output wire[NUMBER_REQUESTS-1:0] llvq_valid,
|
||||
output wire[NUMBER_REQUESTS-1:0][31:0] llvq_res_addr,
|
||||
output wire[NUMBER_REQUESTS-1:0][`BANK_LINE_SIZE_RNG][31:0] llvq_res_data
|
||||
input wire[31:0] snp_req_addr
|
||||
|
||||
);
|
||||
|
||||
@@ -111,8 +106,9 @@ module VX_cache
|
||||
wire [NUMBER_BANKS-1:0][4:0] per_bank_wb_rd;
|
||||
wire [NUMBER_BANKS-1:0][1:0] per_bank_wb_wb;
|
||||
wire [NUMBER_BANKS-1:0][`NW_M1:0] per_bank_wb_warp_num;
|
||||
wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_data;
|
||||
wire [NUMBER_BANKS-1:0][`WORD_SIZE_RNG] per_bank_wb_data;
|
||||
wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_pc;
|
||||
wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_address;
|
||||
|
||||
|
||||
wire dfqq_full;
|
||||
@@ -124,54 +120,15 @@ module VX_cache
|
||||
wire[NUMBER_BANKS-1:0] per_bank_dram_wb_req;
|
||||
wire[NUMBER_BANKS-1:0] per_bank_dram_because_of_snp;
|
||||
wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_wb_req_addr;
|
||||
wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][31:0] per_bank_dram_wb_req_data;
|
||||
wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] per_bank_dram_wb_req_data;
|
||||
|
||||
wire[NUMBER_BANKS-1:0] per_bank_reqq_full;
|
||||
|
||||
|
||||
wire[NUMBER_BANKS-1:0] per_bank_llvq_pop;
|
||||
wire[NUMBER_BANKS-1:0] per_bank_llvq_valid;
|
||||
wire[NUMBER_BANKS-1:0][31:0] per_bank_llvq_res_addr;
|
||||
wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][31:0] per_bank_llvq_res_data;
|
||||
wire [NUMBER_BANKS-1:0][`vx_clog2(NUMBER_REQUESTS)-1:0] per_bank_llvq_res_tid;
|
||||
|
||||
assign delay_req = (|per_bank_reqq_full);
|
||||
|
||||
|
||||
assign dram_fill_accept = (NUMBER_BANKS == 1) ? per_bank_dram_fill_accept[0] : per_bank_dram_fill_accept[dram_fill_rsp_addr[`BANK_SELECT_ADDR_RNG]];
|
||||
|
||||
|
||||
VX_dcache_llv_resp_bank_sel #(
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
.NUMBER_BANKS (NUMBER_BANKS),
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUMBER_REQUESTS (NUMBER_REQUESTS),
|
||||
.STAGE_1_CYCLES (STAGE_1_CYCLES),
|
||||
.REQQ_SIZE (REQQ_SIZE),
|
||||
.MRVQ_SIZE (MRVQ_SIZE),
|
||||
.DFPQ_SIZE (DFPQ_SIZE),
|
||||
.SNRQ_SIZE (SNRQ_SIZE),
|
||||
.CWBQ_SIZE (CWBQ_SIZE),
|
||||
.DWBQ_SIZE (DWBQ_SIZE),
|
||||
.DFQQ_SIZE (DFQQ_SIZE),
|
||||
.LLVQ_SIZE (LLVQ_SIZE),
|
||||
.FILL_INVALIDAOR_SIZE (FILL_INVALIDAOR_SIZE),
|
||||
.SIMULATED_DRAM_LATENCY_CYCLES(SIMULATED_DRAM_LATENCY_CYCLES)
|
||||
)
|
||||
VX_dcache_llv_resp_bank_sel
|
||||
(
|
||||
.per_bank_llvq_pop (per_bank_llvq_pop),
|
||||
.per_bank_llvq_valid (per_bank_llvq_valid),
|
||||
.per_bank_llvq_res_addr(per_bank_llvq_res_addr),
|
||||
.per_bank_llvq_res_data(per_bank_llvq_res_data),
|
||||
.per_bank_llvq_res_tid (per_bank_llvq_res_tid),
|
||||
.llvq_pop (llvq_pop),
|
||||
.llvq_valid (llvq_valid),
|
||||
.llvq_res_addr (llvq_res_addr),
|
||||
.llvq_res_data (llvq_res_data)
|
||||
);
|
||||
|
||||
VX_cache_dram_req_arb #(
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
.BANK_LINE_SIZE_BYTES (BANK_LINE_SIZE_BYTES),
|
||||
@@ -245,6 +202,7 @@ module VX_cache
|
||||
.WORD_SIZE_BYTES (WORD_SIZE_BYTES),
|
||||
.NUMBER_REQUESTS (NUMBER_REQUESTS),
|
||||
.STAGE_1_CYCLES (STAGE_1_CYCLES),
|
||||
.FUNC_ID (FUNC_ID),
|
||||
.REQQ_SIZE (REQQ_SIZE),
|
||||
.MRVQ_SIZE (MRVQ_SIZE),
|
||||
.DFPQ_SIZE (DFPQ_SIZE),
|
||||
@@ -266,6 +224,7 @@ module VX_cache
|
||||
.per_bank_wb_warp_num(per_bank_wb_warp_num),
|
||||
.per_bank_wb_data (per_bank_wb_data),
|
||||
.per_bank_wb_pop (per_bank_wb_pop),
|
||||
.per_bank_wb_address (per_bank_wb_address),
|
||||
|
||||
.core_no_wb_slot (core_no_wb_slot),
|
||||
.core_wb_valid (core_wb_valid),
|
||||
@@ -273,6 +232,7 @@ module VX_cache
|
||||
.core_wb_req_wb (core_wb_req_wb),
|
||||
.core_wb_warp_num (core_wb_warp_num),
|
||||
.core_wb_readdata (core_wb_readdata),
|
||||
.core_wb_address (core_wb_address),
|
||||
.core_wb_pc (core_wb_pc)
|
||||
);
|
||||
|
||||
@@ -281,12 +241,12 @@ module VX_cache
|
||||
for (curr_bank = 0; curr_bank < NUMBER_BANKS; curr_bank=curr_bank+1) begin
|
||||
wire [NUMBER_REQUESTS-1:0] curr_bank_valids;
|
||||
wire [NUMBER_REQUESTS-1:0][31:0] curr_bank_addr;
|
||||
wire [NUMBER_REQUESTS-1:0][31:0] curr_bank_writedata;
|
||||
wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] curr_bank_writedata;
|
||||
wire [4:0] curr_bank_rd;
|
||||
wire [1:0] curr_bank_wb;
|
||||
wire [NUMBER_REQUESTS-1:0][1:0] curr_bank_wb;
|
||||
wire [`NW_M1:0] curr_bank_warp_num;
|
||||
wire [2:0] curr_bank_mem_read;
|
||||
wire [2:0] curr_bank_mem_write;
|
||||
wire [NUMBER_REQUESTS-1:0][2:0] curr_bank_mem_read;
|
||||
wire [NUMBER_REQUESTS-1:0][2:0] curr_bank_mem_write;
|
||||
wire [31:0] curr_bank_pc;
|
||||
|
||||
wire curr_bank_wb_pop;
|
||||
@@ -296,11 +256,12 @@ module VX_cache
|
||||
wire [4:0] curr_bank_wb_rd;
|
||||
wire [1:0] curr_bank_wb_wb;
|
||||
wire [`NW_M1:0] curr_bank_wb_warp_num;
|
||||
wire [31:0] curr_bank_wb_data;
|
||||
wire [`WORD_SIZE_RNG] curr_bank_wb_data;
|
||||
wire [31:0] curr_bank_wb_address;
|
||||
|
||||
wire curr_bank_dram_fill_rsp;
|
||||
wire [31:0] curr_bank_dram_fill_rsp_addr;
|
||||
wire [`BANK_LINE_SIZE_RNG][31:0] curr_bank_dram_fill_rsp_data;
|
||||
wire [`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] curr_bank_dram_fill_rsp_data;
|
||||
wire curr_bank_dram_fill_accept;
|
||||
|
||||
wire curr_bank_dfqq_full;
|
||||
@@ -312,19 +273,13 @@ module VX_cache
|
||||
wire curr_bank_dram_wb_queue_pop;
|
||||
wire curr_bank_dram_wb_req;
|
||||
wire[31:0] curr_bank_dram_wb_req_addr;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] curr_bank_dram_wb_req_data;
|
||||
wire[`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] curr_bank_dram_wb_req_data;
|
||||
|
||||
wire curr_bank_snp_req;
|
||||
wire[31:0] curr_bank_snp_req_addr;
|
||||
|
||||
wire curr_bank_reqq_full;
|
||||
|
||||
|
||||
wire curr_bank_llvq_pop;
|
||||
wire curr_bank_llvq_valid;
|
||||
wire[31:0] curr_bank_llvq_res_addr;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] curr_bank_llvq_res_data;
|
||||
wire[`vx_clog2(NUMBER_REQUESTS)-1:0] curr_bank_llvq_res_tid;
|
||||
|
||||
|
||||
// Core Req
|
||||
@@ -348,6 +303,7 @@ module VX_cache
|
||||
assign per_bank_wb_warp_num[curr_bank] = curr_bank_wb_warp_num;
|
||||
assign per_bank_wb_data [curr_bank] = curr_bank_wb_data;
|
||||
assign per_bank_wb_pc [curr_bank] = curr_bank_wb_pc;
|
||||
assign per_bank_wb_address [curr_bank] = curr_bank_wb_address;
|
||||
|
||||
// Dram fill request
|
||||
assign curr_bank_dfqq_full = dfqq_full;
|
||||
@@ -370,14 +326,6 @@ module VX_cache
|
||||
// Snoop Request
|
||||
assign curr_bank_snp_req = snp_req && (snp_req_addr[`BANK_SELECT_ADDR_RNG] == curr_bank);
|
||||
assign curr_bank_snp_req_addr = snp_req_addr;
|
||||
|
||||
|
||||
// LLVQ
|
||||
assign curr_bank_llvq_pop = per_bank_llvq_pop[curr_bank];
|
||||
assign per_bank_llvq_valid[curr_bank] = curr_bank_llvq_valid;
|
||||
assign per_bank_llvq_res_data[curr_bank] = curr_bank_llvq_res_data;
|
||||
assign per_bank_llvq_res_addr[curr_bank] = curr_bank_llvq_res_addr;
|
||||
assign per_bank_llvq_res_tid[curr_bank] = curr_bank_llvq_res_tid;
|
||||
|
||||
VX_bank #(
|
||||
.CACHE_SIZE_BYTES (CACHE_SIZE_BYTES),
|
||||
@@ -424,6 +372,7 @@ module VX_cache
|
||||
.bank_wb_warp_num (curr_bank_wb_warp_num),
|
||||
.bank_wb_data (curr_bank_wb_data),
|
||||
.bank_wb_pc (curr_bank_wb_pc),
|
||||
.bank_wb_address (curr_bank_wb_address),
|
||||
|
||||
// Dram fill req
|
||||
.dram_fill_req (curr_bank_dram_fill_req),
|
||||
@@ -446,13 +395,8 @@ module VX_cache
|
||||
|
||||
// Snoop Request
|
||||
.snp_req (curr_bank_snp_req),
|
||||
.snp_req_addr (curr_bank_snp_req_addr),
|
||||
.snp_req_addr (curr_bank_snp_req_addr)
|
||||
|
||||
.llvq_pop (curr_bank_llvq_pop),
|
||||
.llvq_valid (curr_bank_llvq_valid),
|
||||
.llvq_res_addr (curr_bank_llvq_res_addr),
|
||||
.llvq_res_data (curr_bank_llvq_res_data),
|
||||
.llvq_res_tid (curr_bank_llvq_res_tid)
|
||||
);
|
||||
|
||||
end
|
||||
|
||||
@@ -4,12 +4,17 @@
|
||||
`include "../VX_define.v"
|
||||
|
||||
|
||||
// data tid rd wb warp_num read write
|
||||
`define MRVQ_METADATA_SIZE (32 + $clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1 + 1) + 3 + 3)
|
||||
// data tid rd wb warp_num read write
|
||||
|
||||
`define REQ_INST_META_SIZE (5 + 2 + (`NW_M1+1) + 3 + 3 + $clog2(NUMBER_REQUESTS))
|
||||
|
||||
`define vx_clog2(value) $clog2(value)
|
||||
`define vx_clog2(value) ((value == 1) ? 1 : $clog2(value))
|
||||
|
||||
|
||||
`define MRVQ_METADATA_SIZE (`WORD_SIZE + `vx_clog2(NUMBER_REQUESTS) + 5 + 2 + (`NW_M1 + 1) + 3 + 3)
|
||||
|
||||
// 5 + 2 + 4 + 3 + 3 + 1
|
||||
`define REQ_INST_META_SIZE (5 + 2 + (`NW_M1+1) + 3 + 3 + `vx_clog2(NUMBER_REQUESTS))
|
||||
|
||||
// `define vx_clog2_h(value, x) (value == (1 << x)) ? (x)
|
||||
|
||||
// `define vx_clog2(value) (value == 0 ) ? 0 : \
|
||||
@@ -46,6 +51,9 @@
|
||||
// `vx_clog2_h(value, 31) : \
|
||||
// 0
|
||||
|
||||
`define WORD_SIZE (8*WORD_SIZE_BYTES)
|
||||
`define WORD_SIZE_RNG (`WORD_SIZE)-1:0
|
||||
|
||||
// 128
|
||||
`define BANK_SIZE_BYTES CACHE_SIZE_BYTES/NUMBER_BANKS
|
||||
|
||||
@@ -65,7 +73,7 @@
|
||||
`define OFFSET_SIZE_RNG `OFFSET_SIZE_END:0
|
||||
|
||||
// 2
|
||||
`define WORD_SELECT_NUM_BITS ($clog2(`BANK_LINE_SIZE_WORDS))
|
||||
`define WORD_SELECT_NUM_BITS (`vx_clog2(`BANK_LINE_SIZE_WORDS))
|
||||
// 2
|
||||
`define WORD_SELECT_SIZE_END (`WORD_SELECT_NUM_BITS)
|
||||
// 2
|
||||
@@ -77,7 +85,7 @@
|
||||
`define WORD_SELECT_SIZE_RNG `WORD_SELECT_SIZE_END-1:0
|
||||
|
||||
// 3
|
||||
`define BANK_SELECT_NUM_BITS ($clog2(NUMBER_BANKS))
|
||||
`define BANK_SELECT_NUM_BITS (`vx_clog2(NUMBER_BANKS))
|
||||
// 3
|
||||
`define BANK_SELECT_SIZE_END (`BANK_SELECT_NUM_BITS)
|
||||
// 4
|
||||
@@ -90,7 +98,7 @@
|
||||
`define BANK_SELECT_SIZE_RNG `BANK_SELECT_SIZE_END-1:0
|
||||
|
||||
// 3
|
||||
`define LINE_SELECT_NUM_BITS ($clog2(`BANK_LINE_COUNT))
|
||||
`define LINE_SELECT_NUM_BITS (`vx_clog2(`BANK_LINE_COUNT))
|
||||
// 3
|
||||
`define LINE_SELECT_SIZE_END (`LINE_SELECT_NUM_BITS)
|
||||
// 7
|
||||
|
||||
@@ -15,6 +15,8 @@ module VX_cache_core_req_bank_sel
|
||||
parameter NUMBER_REQUESTS = 2,
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
|
||||
@@ -58,7 +58,7 @@ module VX_cache_dram_req_arb
|
||||
output wire[NUMBER_BANKS-1:0] per_bank_dram_wb_queue_pop,
|
||||
input wire[NUMBER_BANKS-1:0] per_bank_dram_wb_req,
|
||||
input wire[NUMBER_BANKS-1:0][31:0] per_bank_dram_wb_req_addr,
|
||||
input wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][31:0] per_bank_dram_wb_req_data,
|
||||
input wire[NUMBER_BANKS-1:0][`BANK_LINE_SIZE_RNG][`WORD_SIZE-1:0] per_bank_dram_wb_req_data,
|
||||
input wire[NUMBER_BANKS-1:0] per_bank_dram_because_of_snp,
|
||||
|
||||
// real Dram request
|
||||
@@ -67,7 +67,7 @@ module VX_cache_dram_req_arb
|
||||
output wire dram_req_read,
|
||||
output wire [31:0] dram_req_addr,
|
||||
output wire [31:0] dram_req_size,
|
||||
output wire [`BANK_LINE_SIZE_RNG][31:0] dram_req_data,
|
||||
output wire [`IBANK_LINE_SIZE_RNG][31:0] dram_req_data,
|
||||
output wire dram_req_because_of_wb
|
||||
|
||||
);
|
||||
@@ -109,7 +109,7 @@ module VX_cache_dram_req_arb
|
||||
assign dram_req_read = dfqq_req && !dwb_valid;
|
||||
assign dram_req_addr = (dwb_valid ? per_bank_dram_wb_req_addr[dwb_bank] : dfqq_req_addr) & `BASE_ADDR_MASK;
|
||||
assign dram_req_size = BANK_LINE_SIZE_BYTES;
|
||||
assign dram_req_data = dwb_valid ? per_bank_dram_wb_req_data[dwb_bank] : 0;
|
||||
assign {dram_req_data} = dwb_valid ? {per_bank_dram_wb_req_data[dwb_bank] }: 0;
|
||||
assign dram_req_because_of_wb = dwb_valid ? per_bank_dram_because_of_snp[dwb_bank] : 0;
|
||||
|
||||
endmodule
|
||||
@@ -52,7 +52,7 @@ module VX_cache_miss_resrv
|
||||
// Miss enqueue
|
||||
input wire miss_add,
|
||||
input wire[31:0] miss_add_addr,
|
||||
input wire[31:0] miss_add_data,
|
||||
input wire[`WORD_SIZE_RNG] miss_add_data,
|
||||
input wire[`vx_clog2(NUMBER_REQUESTS)-1:0] miss_add_tid,
|
||||
input wire[4:0] miss_add_rd,
|
||||
input wire[1:0] miss_add_wb,
|
||||
@@ -70,7 +70,7 @@ module VX_cache_miss_resrv
|
||||
input wire miss_resrv_pop,
|
||||
output wire miss_resrv_valid_st0,
|
||||
output wire[31:0] miss_resrv_addr_st0,
|
||||
output wire[31:0] miss_resrv_data_st0,
|
||||
output wire[`WORD_SIZE_RNG] miss_resrv_data_st0,
|
||||
output wire[`vx_clog2(NUMBER_REQUESTS)-1:0] miss_resrv_tid_st0,
|
||||
output wire[4:0] miss_resrv_rd_st0,
|
||||
output wire[1:0] miss_resrv_wb_st0,
|
||||
|
||||
@@ -52,12 +52,12 @@ module VX_cache_req_queue
|
||||
input wire reqq_push,
|
||||
input wire [NUMBER_REQUESTS-1:0] bank_valids,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] bank_addr,
|
||||
input wire [NUMBER_REQUESTS-1:0][31:0] bank_writedata,
|
||||
input wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] bank_writedata,
|
||||
input wire [4:0] bank_rd,
|
||||
input wire [1:0] bank_wb,
|
||||
input wire [NUMBER_REQUESTS-1:0][1:0] bank_wb,
|
||||
input wire [`NW_M1:0] bank_warp_num,
|
||||
input wire [2:0] bank_mem_read,
|
||||
input wire [2:0] bank_mem_write,
|
||||
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_read,
|
||||
input wire [NUMBER_REQUESTS-1:0][2:0] bank_mem_write,
|
||||
input wire [31:0] bank_pc,
|
||||
|
||||
// Dequeue Data
|
||||
@@ -65,7 +65,7 @@ module VX_cache_req_queue
|
||||
output wire reqq_req_st0,
|
||||
output wire [`vx_clog2(NUMBER_REQUESTS)-1:0] reqq_req_tid_st0,
|
||||
output wire [31:0] reqq_req_addr_st0,
|
||||
output wire [31:0] reqq_req_writedata_st0,
|
||||
output wire [`WORD_SIZE_RNG] reqq_req_writedata_st0,
|
||||
output wire [4:0] reqq_req_rd_st0,
|
||||
output wire [1:0] reqq_req_wb_st0,
|
||||
output wire [`NW_M1:0] reqq_req_warp_num_st0,
|
||||
@@ -80,34 +80,34 @@ module VX_cache_req_queue
|
||||
|
||||
wire [NUMBER_REQUESTS-1:0] out_per_valids;
|
||||
wire [NUMBER_REQUESTS-1:0][31:0] out_per_addr;
|
||||
wire [NUMBER_REQUESTS-1:0][31:0] out_per_writedata;
|
||||
wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] out_per_writedata;
|
||||
wire [4:0] out_per_rd;
|
||||
wire [1:0] out_per_wb;
|
||||
wire [NUMBER_REQUESTS-1:0][1:0] out_per_wb;
|
||||
wire [`NW_M1:0] out_per_warp_num;
|
||||
wire [2:0] out_per_mem_read;
|
||||
wire [2:0] out_per_mem_write;
|
||||
wire [NUMBER_REQUESTS-1:0][2:0] out_per_mem_read;
|
||||
wire [NUMBER_REQUESTS-1:0][2:0] out_per_mem_write;
|
||||
wire [31:0] out_per_pc;
|
||||
|
||||
|
||||
reg [NUMBER_REQUESTS-1:0] use_per_valids;
|
||||
reg [NUMBER_REQUESTS-1:0][31:0] use_per_addr;
|
||||
reg [NUMBER_REQUESTS-1:0][31:0] use_per_writedata;
|
||||
reg [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] use_per_writedata;
|
||||
reg [4:0] use_per_rd;
|
||||
reg [1:0] use_per_wb;
|
||||
reg [NUMBER_REQUESTS-1:0][1:0] use_per_wb;
|
||||
reg [31:0] use_per_pc;
|
||||
reg [`NW_M1:0] use_per_warp_num;
|
||||
reg [2:0] use_per_mem_read;
|
||||
reg [2:0] use_per_mem_write;
|
||||
reg [NUMBER_REQUESTS-1:0][2:0] use_per_mem_read;
|
||||
reg [NUMBER_REQUESTS-1:0][2:0] use_per_mem_write;
|
||||
|
||||
|
||||
wire [NUMBER_REQUESTS-1:0] qual_valids;
|
||||
wire [NUMBER_REQUESTS-1:0][31:0] qual_addr;
|
||||
wire [NUMBER_REQUESTS-1:0][31:0] qual_writedata;
|
||||
wire [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] qual_writedata;
|
||||
wire [4:0] qual_rd;
|
||||
wire [1:0] qual_wb;
|
||||
wire [NUMBER_REQUESTS-1:0][1:0] qual_wb;
|
||||
wire [`NW_M1:0] qual_warp_num;
|
||||
wire [2:0] qual_mem_read;
|
||||
wire [2:0] qual_mem_write;
|
||||
wire [NUMBER_REQUESTS-1:0][2:0] qual_mem_read;
|
||||
wire [NUMBER_REQUESTS-1:0][2:0] qual_mem_write;
|
||||
wire [31:0] qual_pc;
|
||||
|
||||
wire[NUMBER_REQUESTS-1:0] updated_valids;
|
||||
@@ -120,7 +120,7 @@ module VX_cache_req_queue
|
||||
wire push_qual = reqq_push && !reqq_full;
|
||||
wire pop_qual = reqq_pop && use_empty && !out_empty;
|
||||
|
||||
VX_generic_queue_ll #(.DATAW( (NUMBER_REQUESTS * (1+32+32)) + 5 + 2 + (`NW_M1+1) + 3 + 3 + 32 ), .SIZE(REQQ_SIZE)) reqq_queue(
|
||||
VX_generic_queue_ll #(.DATAW( (NUMBER_REQUESTS * (1+32+`WORD_SIZE)) + 5 + (NUMBER_REQUESTS*2) + (`NW_M1+1) + (NUMBER_REQUESTS * (3 + 3)) + 32 ), .SIZE(REQQ_SIZE)) reqq_queue(
|
||||
.clk (clk),
|
||||
.reset (reset),
|
||||
.push (push_qual),
|
||||
@@ -158,10 +158,10 @@ module VX_cache_req_queue
|
||||
assign reqq_req_addr_st0 = qual_addr [qual_request_index];
|
||||
assign reqq_req_writedata_st0 = qual_writedata[qual_request_index];
|
||||
assign reqq_req_rd_st0 = qual_rd;
|
||||
assign reqq_req_wb_st0 = qual_wb;
|
||||
assign reqq_req_wb_st0 = qual_wb[qual_request_index];
|
||||
assign reqq_req_warp_num_st0 = qual_warp_num;
|
||||
assign reqq_req_mem_read_st0 = qual_mem_read;
|
||||
assign reqq_req_mem_write_st0 = qual_mem_write;
|
||||
assign reqq_req_mem_read_st0 = qual_mem_read [qual_request_index];
|
||||
assign reqq_req_mem_write_st0 = qual_mem_write[qual_request_index];
|
||||
assign reqq_req_pc_st0 = qual_pc;
|
||||
|
||||
assign updated_valids = qual_valids & (~(1 << qual_request_index));
|
||||
|
||||
@@ -14,6 +14,8 @@ module VX_cache_wb_sel_merge
|
||||
parameter NUMBER_REQUESTS = 2,
|
||||
// Number of cycles to complete stage 1 (read from memory)
|
||||
parameter STAGE_1_CYCLES = 2,
|
||||
// Function ID, {Dcache=0, Icache=1, Sharedmemory=2}
|
||||
parameter FUNC_ID = 0,
|
||||
|
||||
// Queues feeding into banks Knobs {1, 2, 4, 8, ...}
|
||||
|
||||
@@ -52,19 +54,21 @@ module VX_cache_wb_sel_merge
|
||||
input wire [NUMBER_BANKS-1:0][4:0] per_bank_wb_rd,
|
||||
input wire [NUMBER_BANKS-1:0][1:0] per_bank_wb_wb,
|
||||
input wire [NUMBER_BANKS-1:0][`NW_M1:0] per_bank_wb_warp_num,
|
||||
input wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_data,
|
||||
input wire [NUMBER_BANKS-1:0][`WORD_SIZE_RNG] per_bank_wb_data,
|
||||
input wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_pc,
|
||||
input wire [NUMBER_BANKS-1:0][31:0] per_bank_wb_address,
|
||||
output wire [NUMBER_BANKS-1:0] per_bank_wb_pop,
|
||||
|
||||
|
||||
// Core Writeback
|
||||
input wire core_no_wb_slot,
|
||||
output reg [NUMBER_REQUESTS-1:0] core_wb_valid,
|
||||
output reg [NUMBER_REQUESTS-1:0][31:0] core_wb_readdata,
|
||||
output reg [NUMBER_REQUESTS-1:0][`WORD_SIZE_RNG] core_wb_readdata,
|
||||
output reg [NUMBER_REQUESTS-1:0][31:0] core_wb_pc,
|
||||
output wire [4:0] core_wb_req_rd,
|
||||
output wire [1:0] core_wb_req_wb,
|
||||
output wire [`NW_M1:0] core_wb_warp_num
|
||||
output wire [`NW_M1:0] core_wb_warp_num,
|
||||
output reg [NUMBER_REQUESTS-1:0][31:0] core_wb_address
|
||||
|
||||
);
|
||||
|
||||
@@ -99,11 +103,13 @@ module VX_cache_wb_sel_merge
|
||||
core_wb_valid = 0;
|
||||
core_wb_readdata = 0;
|
||||
core_wb_pc = 0;
|
||||
core_wb_address = 0;
|
||||
for (this_bank = 0; this_bank < NUMBER_BANKS; this_bank = this_bank + 1) begin
|
||||
if (found_bank && (per_bank_wb_valid[this_bank]) && (per_bank_wb_rd[this_bank] == per_bank_wb_rd[main_bank_index]) && (per_bank_wb_warp_num[this_bank] == per_bank_wb_warp_num[main_bank_index])) begin
|
||||
if (((FUNC_ID == `LLFUNC_ID) && found_bank && per_bank_wb_valid[this_bank]) || (found_bank && (per_bank_wb_valid[this_bank]) && (per_bank_wb_rd[this_bank] == per_bank_wb_rd[main_bank_index]) && (per_bank_wb_warp_num[this_bank] == per_bank_wb_warp_num[main_bank_index]))) begin
|
||||
core_wb_valid[per_bank_wb_tid[this_bank]] = 1;
|
||||
core_wb_readdata[per_bank_wb_tid[this_bank]] = per_bank_wb_data[this_bank];
|
||||
core_wb_pc[per_bank_wb_tid[this_bank]] = per_bank_wb_pc[this_bank];
|
||||
core_wb_address[per_bank_wb_tid[this_bank]] = per_bank_wb_address[this_bank];
|
||||
per_bank_wb_pop_unqual[this_bank] = 1;
|
||||
end else begin
|
||||
per_bank_wb_pop_unqual[this_bank] = 0;
|
||||
|
||||
@@ -58,13 +58,13 @@ module VX_tag_data_access
|
||||
input wire valid_req_st1e,
|
||||
input wire writefill_st1e,
|
||||
input wire[31:0] writeaddr_st1e,
|
||||
input wire[31:0] writeword_st1e,
|
||||
input wire[`BANK_LINE_SIZE_RNG][31:0] writedata_st1e,
|
||||
input wire[`WORD_SIZE_RNG] writeword_st1e,
|
||||
input wire[`DBANK_LINE_SIZE_RNG][31:0] writedata_st1e,
|
||||
input wire[2:0] mem_write_st1e,
|
||||
input wire[2:0] mem_read_st1e,
|
||||
|
||||
output wire[31:0] readword_st1e,
|
||||
output wire[`BANK_LINE_SIZE_RNG][31:0] readdata_st1e,
|
||||
output wire[`WORD_SIZE_RNG] readword_st1e,
|
||||
output wire[`DBANK_LINE_SIZE_RNG][31:0] readdata_st1e,
|
||||
output wire[`TAG_SELECT_SIZE_RNG] readtag_st1e,
|
||||
output wire miss_st1e,
|
||||
output wire dirty_st1e,
|
||||
@@ -73,25 +73,25 @@ module VX_tag_data_access
|
||||
);
|
||||
|
||||
|
||||
reg[`BANK_LINE_SIZE_RNG][31:0] readdata_st[STAGE_1_CYCLES-2:0];
|
||||
reg[`DBANK_LINE_SIZE_RNG][31:0] readdata_st[STAGE_1_CYCLES-2:0];
|
||||
|
||||
reg read_valid_st1c[STAGE_1_CYCLES-2:0];
|
||||
reg read_dirty_st1c[STAGE_1_CYCLES-2:0];
|
||||
reg[`TAG_SELECT_SIZE_RNG] read_tag_st1c [STAGE_1_CYCLES-2:0];
|
||||
reg[`BANK_LINE_SIZE_RNG][31:0] read_data_st1c [STAGE_1_CYCLES-2:0];
|
||||
reg[`DBANK_LINE_SIZE_RNG][31:0] read_data_st1c [STAGE_1_CYCLES-2:0];
|
||||
|
||||
|
||||
wire qual_read_valid_st1;
|
||||
wire qual_read_dirty_st1;
|
||||
wire[`TAG_SELECT_SIZE_RNG] qual_read_tag_st1;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] qual_read_data_st1;
|
||||
wire[`DBANK_LINE_SIZE_RNG][31:0] qual_read_data_st1;
|
||||
|
||||
wire use_read_valid_st1e;
|
||||
wire use_read_dirty_st1e;
|
||||
wire[`TAG_SELECT_SIZE_RNG] use_read_tag_st1e;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] use_read_data_st1e;
|
||||
wire[`BANK_LINE_SIZE_RNG][3:0] use_write_enable;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] use_write_data;
|
||||
wire[`DBANK_LINE_SIZE_RNG][31:0] use_read_data_st1e;
|
||||
wire[`DBANK_LINE_SIZE_RNG][3:0] use_write_enable;
|
||||
wire[`DBANK_LINE_SIZE_RNG][31:0] use_write_data;
|
||||
|
||||
|
||||
wire fill_sent;
|
||||
@@ -134,7 +134,7 @@ module VX_tag_data_access
|
||||
.fill_sent (fill_sent)
|
||||
);
|
||||
|
||||
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`BANK_LINE_SIZE_WORDS*32) )) s0_1_c0 (
|
||||
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_SIZE_WORDS*32) )) s0_1_c0 (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall),
|
||||
@@ -146,7 +146,7 @@ module VX_tag_data_access
|
||||
genvar curr_stage;
|
||||
generate
|
||||
for (curr_stage = 1; curr_stage < STAGE_1_CYCLES-2; curr_stage = curr_stage + 1) begin
|
||||
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`BANK_LINE_SIZE_WORDS*32) )) s0_1_cc (
|
||||
VX_generic_register #(.N( 1 + 1 + `TAG_SELECT_NUM_BITS + (`DBANK_LINE_SIZE_WORDS*32) )) s0_1_cc (
|
||||
.clk (clk),
|
||||
.reset(reset),
|
||||
.stall(stall),
|
||||
@@ -163,7 +163,7 @@ module VX_tag_data_access
|
||||
assign use_read_tag_st1e = (FUNC_ID == `SFUNC_ID) ? writeaddr_st1e[`TAG_SELECT_ADDR_RNG] : read_tag_st1c [STAGE_1_CYCLES-2]; // Tag is always the same in SM
|
||||
|
||||
genvar curr_w;
|
||||
for (curr_w = 0; curr_w < `BANK_LINE_SIZE_WORDS; curr_w = curr_w+1) assign use_read_data_st1e[curr_w][31:0] = read_data_st1c[STAGE_1_CYCLES-2][curr_w][31:0];
|
||||
for (curr_w = 0; curr_w < `DBANK_LINE_SIZE_WORDS; curr_w = curr_w+1) assign use_read_data_st1e[curr_w][31:0] = read_data_st1c[STAGE_1_CYCLES-2][curr_w][31:0];
|
||||
// assign use_read_data_st1e = read_data_st1c [STAGE_1_CYCLES-2];
|
||||
|
||||
/////////////////////// LOAD LOGIC ///////////////////
|
||||
@@ -202,14 +202,14 @@ module VX_tag_data_access
|
||||
wire[31:0] lw_data = (data_unQual);
|
||||
|
||||
|
||||
wire[31:0] sw_data = writeword_st1e;
|
||||
wire[31:0] sw_data = writeword_st1e[31:0];
|
||||
|
||||
wire[31:0] sb_data = b1 ? {{16{1'b0}}, writeword_st1e[7:0], { 8{1'b0}}} :
|
||||
b2 ? {{ 8{1'b0}}, writeword_st1e[7:0], {16{1'b0}}} :
|
||||
b3 ? {{ 0{1'b0}}, writeword_st1e[7:0], {24{1'b0}}} :
|
||||
writeword_st1e;
|
||||
writeword_st1e[31:0];
|
||||
|
||||
wire[31:0] sh_data = b2 ? {writeword_st1e[15:0], {16{1'b0}}} : writeword_st1e;
|
||||
wire[31:0] sh_data = b2 ? {writeword_st1e[15:0], {16{1'b0}}} : writeword_st1e[31:0];
|
||||
|
||||
|
||||
|
||||
@@ -236,20 +236,24 @@ module VX_tag_data_access
|
||||
wire should_write = (sw || sb || sh) && valid_req_st1e && use_read_valid_st1e && !miss_st1e;
|
||||
wire force_write = writefill_st1e && valid_req_st1e && (!use_read_valid_st1e || (use_read_valid_st1e && !miss_st1e));
|
||||
|
||||
wire[`BANK_LINE_SIZE_RNG][3:0] we;
|
||||
wire[`BANK_LINE_SIZE_RNG][31:0] data_write;
|
||||
wire[`DBANK_LINE_SIZE_RNG][3:0] we;
|
||||
wire[`DBANK_LINE_SIZE_RNG][31:0] data_write;
|
||||
genvar g;
|
||||
generate
|
||||
for (g = 0; g < `BANK_LINE_SIZE_WORDS; g = g + 1) begin : write_enables
|
||||
wire normal_write = (block_offset == g) && should_write && !writefill_st1e;
|
||||
for (g = 0; g < `DBANK_LINE_SIZE_WORDS; g = g + 1) begin : write_enables
|
||||
wire normal_write = (block_offset == g[`WORD_SELECT_SIZE_RNG]) && should_write && !writefill_st1e;
|
||||
|
||||
assign we[g] = (force_write) ? 4'b1111 :
|
||||
(normal_write && (FUNC_ID == `LLFUNC_ID)) ? 4'b1111 :
|
||||
(normal_write && sw) ? 4'b1111 :
|
||||
(normal_write && sb) ? sb_mask :
|
||||
(normal_write && sh) ? sh_mask :
|
||||
4'b0000;
|
||||
|
||||
assign data_write[g] = force_write ? writedata_st1e[g] : use_write_dat;
|
||||
if (!(FUNC_ID == `LLFUNC_ID)) assign data_write[g] = force_write ? writedata_st1e[g] : use_write_dat;
|
||||
end
|
||||
if ((FUNC_ID == `LLFUNC_ID)) begin
|
||||
assign data_write = force_write ? writedata_st1e : writeword_st1e;
|
||||
end
|
||||
endgenerate
|
||||
|
||||
@@ -257,8 +261,12 @@ module VX_tag_data_access
|
||||
assign use_write_data = data_write;
|
||||
|
||||
///////////////////////
|
||||
|
||||
assign readword_st1e = data_Qual;
|
||||
if (FUNC_ID == `LLFUNC_ID) begin
|
||||
assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-2];
|
||||
end else begin
|
||||
assign readword_st1e = data_Qual;
|
||||
end
|
||||
|
||||
assign miss_st1e = ((valid_req_st1e || is_snp_st1e) && !use_read_valid_st1e) || (valid_req_st1e && use_read_valid_st1e && !writefill_st1e && (writeaddr_st1e[`TAG_SELECT_ADDR_RNG] != use_read_tag_st1e));
|
||||
assign dirty_st1e = valid_req_st1e && use_read_valid_st1e && use_read_dirty_st1e;
|
||||
assign readdata_st1e = use_read_data_st1e;
|
||||
|
||||
@@ -54,18 +54,18 @@ module VX_tag_data_structure
|
||||
output wire read_valid,
|
||||
output wire read_dirty,
|
||||
output wire[`TAG_SELECT_SIZE_RNG] read_tag,
|
||||
output wire[`BANK_LINE_SIZE_RNG][31:0] read_data,
|
||||
output wire[`DBANK_LINE_SIZE_RNG][31:0] read_data,
|
||||
|
||||
input wire invalidate,
|
||||
input wire[`BANK_LINE_SIZE_RNG][3:0] write_enable,
|
||||
input wire[`DBANK_LINE_SIZE_RNG][3:0] write_enable,
|
||||
input wire write_fill,
|
||||
input wire[31:0] write_addr,
|
||||
input wire[`BANK_LINE_SIZE_RNG][31:0] write_data,
|
||||
input wire[`DBANK_LINE_SIZE_RNG][31:0] write_data,
|
||||
input wire fill_sent
|
||||
|
||||
);
|
||||
|
||||
reg[`BANK_LINE_SIZE_RNG][3:0][7:0] data [`BANK_LINE_COUNT-1:0];
|
||||
reg[`DBANK_LINE_SIZE_RNG][3:0][7:0] data [`BANK_LINE_COUNT-1:0];
|
||||
reg[`TAG_SELECT_SIZE_RNG] tag [`BANK_LINE_COUNT-1:0];
|
||||
reg valid[`BANK_LINE_COUNT-1:0];
|
||||
reg dirty[`BANK_LINE_COUNT-1:0];
|
||||
@@ -98,7 +98,7 @@ module VX_tag_data_structure
|
||||
valid[write_addr[`LINE_SELECT_ADDR_RNG]] <= 0;
|
||||
end
|
||||
|
||||
for (f = 0; f < `BANK_LINE_SIZE_WORDS; f = f + 1) begin
|
||||
for (f = 0; f < `DBANK_LINE_SIZE_WORDS; f = f + 1) begin
|
||||
if (write_enable[f][0]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][0] <= write_data[f][7 :0 ];
|
||||
if (write_enable[f][1]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][1] <= write_data[f][15:8 ];
|
||||
if (write_enable[f][2]) data[write_addr[`LINE_SELECT_ADDR_RNG]][f][2] <= write_data[f][23:16];
|
||||
|
||||
Reference in New Issue
Block a user